RECTIFIABILITY OF SELF-CONTRACTED CURVES
IN THE EUCLIDEAN SPACE AND APPLICATIONS
A. DANIILIDIS, G. DAVID, E. DURAND-CARTAGENA, A. LEMENANT
Abstract. It is hereby established that, in Euclidean spaces of finite dimension, bounded selfcontracted curves have finite length. This extends the main result of [6] concerning continuous
planar self-contracted curves to any dimension, and dispenses entirely with the continuity requirement. The proof borrows heavily from a geometric idea of [13] employed for the study
of regular enough curves, and can be seen as a nonsmooth adaptation of the latter, albeit a
nontrivial one. Applications to continuous and discrete dynamical systems are discussed: continuous self-contracted curves appear as generalized solutions of nonsmooth convex foliation
systems, recovering a hidden regularity after reparameterization, as consequence of our main
result. In the discrete case, proximal sequences (obtained through implicit discretization of a
gradient system) give rise to polygonal self-contracted curves. This yields a straightforward
proof for the convergence of the exact proximal algorithm, under any choice of parameters.
Contents
1.
Introduction
2
1.1.
Motivation and state-of-the-art
2
1.2.
Contributions of this work
3
1.3.
Acknowledgements
4
2.
Notation and preliminaries
4
2.1.
Self-expanded curves
5
2.2.
Self-contracted versus self-expanded curves
6
3.
Rectifiability of self-contracted curves
3.1.
Proof of the main result
3.2.
Arc-length reparameterization
4.
7
7
14
Applications
15
4.1.
Orbits of convex foliations
15
4.2.
Polygonal approximations of smooth strongly self-contracted curves
19
4.3.
Convergence of the proximal algorithm
21
References
23
2000 Mathematics Subject Classification. Primary 53A04 ; Secondary 37N40, 49J52, 49J53, 52A10, 65K10.
Key words and phrases. Self-contracted curve, rectifiable curve, convex foliation, secant, self-expanded curve,
proximal algorithm.
1
2
A. DANIILIDIS, G. DAVID, E. DURAND-CARTAGENA, A. LEMENANT
1. Introduction
1.1. Motivation and state-of-the-art. Self-contracted curves were introduced in [6, Definition 1.2.] to provide a unified framework for the study of convex and quasiconvex gradient
dynamical systems. Given a possibly unbounded interval I of R, a map γ : I → Rn is called
self-contracted curve, if for every [a, b] ⊂ I, the real-valued function
t ∈ [a, b] 7→ d(γ(t), γ(b))
is non-increasing. This notion is purely metric and does not require any prior smoothness/continuity
assumption on γ.
So far, self-contracted curves are considered in a Euclidean framework. In particular, given a
smooth function f : Rn → R, any solution γ of the gradient system
(
γ 0 (t) = −∇f (γ(t))
t > 0,
(1.1)
γ(0) = x0 ∈ Rn
is a (smooth) self-contracted curve, provided that f is quasiconvex, that is, its sublevel sets
[f ≤ β] := {x ∈ Rn : f (x) ≤ β}
(β ∈ R)
are convex subsets of Rn [6, Proposition 6.2]. Self-contracted curves also appear in subgradient
systems, defined by a (nonsmooth) convex function f , see [6, Proposition 6.4]. In this case, the
first equation in (1.1) becomes the following differential inclusion
γ 0 (t) ∈ −∂f (γ(t))
a.e.,
(1.2)
where a.e. stands for “almost everywhere” and the solutions are absolutely continuous curves
(see [2] for a general theory).
A central question of the asymptotic theory of a general gradient dynamical system of the form
(1.1) is whether or not bounded orbits are of finite length, which if true, yields in particular their
convergence. This property fails for C ∞ smooth functions ([15, p. 12]), but holds for analytic
gradient systems ([11]), or more generally, for systems defined by an o-minimal (tame) function
([9]). In these cases, a concrete estimation of the length of (sub)gradient curves is obtained in
terms of a so-called Lojasiewicz type inequality, intrinsically linked to the potential f , see [1] for
a survey, which also includes extensions of the theory to subgradient systems. Notice however
that convex functions fail to satisfy such inequality, see [1, Section 4.3] for a counterexample.
In [13] a certain class of Lipschitz curves has been introduced (with no specific name) to
capture the behaviour of orbits of quasiconvex potentials. Unlike self-contracted curves, this
notion makes sense only in Euclidean spaces (it uses orthogonality) and requires the curve to be
Lipschitz continuous. The main result of [13] asserts that the length of such curves is bounded by
the mean width of their convex hull, a fortiori, by the mean width of any convex set containing
the curve. Recently, the authors of [7] extended the result of [13] to 2-dimensional surfaces
of constant curvature, naming these curves as (G)-orbits. We shall prefer to call these curves
self-expanding, see Definition 2.2 below. The choice of this terminology will become clear in
Section 2.2.
Formally the result, as announced in [13, IX], applies only under the prior requirement that
the length of the curve is finite, since the proof is given for self-expanded curves parameterized by
the arc-length parametrization in a compact interval. This restriction is removed in Section 2.1
via a simple continuity argument (see Corollary 2.4). As a result, the smooth orbits of (1.1)
for a quasiconvex potential, as well as the absolutely continuous orbits of (1.2) for a convex one
RECTIFIABILITY OF SELF-CONTRACTED CURVES
IN THE EUCLIDEAN SPACE AND APPLICATIONS 3
have finite length. In both cases the bound depends only on the diameter of the initial sublevel
set.
The work [13] was unknown to the authors of [6], who tackled the same problem in terms
of the aforementioned notion of self-contracted curve. The definition of self-contracted curve
does not require any regularity neither on the space nor on the curve. In particular, such
curves need not be continuous, and differentiability may a priori fail at each point. The main
result of [6] shows that bounded continuous planar self-contracted curves have finite length ([6,
Theorem 1.3]). This has been used to deduce that in R2 , smooth orbits of quasiconvex systems
(respectively, absolutely continuous orbits of nonsmooth convex systems) have finite length. As
we saw before, this conclusion essentially derives from [13, IX] for any dimension (see comments
above). Notice however that the main result of [6] cannot be deduced from [13], but can only
be compared in retrospect, once rectifiability is established.
On the other hand, not completely surprisingly, Lipschitz continuous self-contracted curves
and self-expanded curves turn out to be intimately related and can be obtained one from the
other by means of an adequate reparameterization, reversing orientations (see Lemma 2.8). Recall however that both rectifiability and Lipschitz continuity of the curve are prior requirements
for the definition of a self-expanded curve, while they are neither requirements nor obvious
consequences of the definition of a self-contracted curve.
1.2. Contributions of this work. In this work we establish rectifiability of any self-contracted
curve in Rn , by extending the result of [6] to any dimension, and to possibly discontinuous curves.
This is done by adapting the geometrical idea of [13] to the class of self-contracted curves. This
nonsmooth adaptation is natural but quite involved. Nonsmooth variations of the mean width of
the closed convex hull of the curve are again used to control the increase of its length, but no prior
continuity on the parametrization is required and rectifiability is now part of the conclusions.
Namely, setting Γ = γ(I) (the image of the curve in Rn ) and denoting by `(γ) its length, we
establish the following result (see Theorem 3.3) :
• Every self-contracted curve is rectifiable and satisfies the relation `(γ) ≤ C diam(Γ),
where C > 0 depends only on the dimension.
In case of continuous curves, the above result allows to consider a Lipschitz reparameterization
defined by the arc-length, see details in Section 3.2. This leads to the following conclusion:
• If γ is a continuous self-contracted curve and Γ = γ(I) is bounded, then Γ is also the
image of some (Lipschitz) self-expanded curve.
In particular, the sets of all possible images of continuous self-contracted curves and of selfexpanded curves coincide. Still, the set of images of all self-contracted curves is much larger (its
elements are not connected in general).
In the last two sections, two new applications of self-contracted curves are considered. In
Section 4.1 we broaden the framework of dynamical systems to encompass nonsmooth convex
foliation systems, with merely continuous generalized orbits. Limits of backward secants remedy
the absence of differentiability, leading to a consistent notion of generalized solution in the sense
of nonsmooth analysis (Definition 4.4). In Theorem 4.6 we show that these generalized solutions
are self-contracted curves, thus of finite length; in view of the aforementioned result, they can
also be obtained as “classical” solutions through an adequate Lipschitz reparameterization.
On the other hand, C 1 smooth convex foliation orbits enjoy a stronger property, the so-called
strong self-contractedness, see Definition 4.8 and Corollary 4.11. Concerning this latter class,
4
A. DANIILIDIS, G. DAVID, E. DURAND-CARTAGENA, A. LEMENANT
we establish in Section 4.2 the following approximation result, with respect to the Hausdorff
distance, see Proposition 4.13.
• Every C 1 -smooth strongly self-contracted curve is a limit of polygonal self-contracted
curves.
Finally, in Section 4.3 we provide an elegant application of the notion of self-contracted curve
in a different framework, that of discrete systems. In particular we establish the following result
(Theorem 4.17) :
• Let f be any convex function, bounded from below. Then the exact proximal algorithm
gives rise to a self-contracted polygonal curve.
In view of our main result, we obtain a straightforward proof of the convergence of the
proximal algorithm. The bound over the length of the polygonal curve yields a sharp estimation
on the rate of convergence, which appears to be entirely new. Notice that the convergence result
is independent of the choice of the parameters.
1.3. Acknowledgements. The first author acknowledges support of the grant MTM201129064-C01 (Spain) and thanks Jerome Bolte and Joel Benoist for useful discussions. The third
author is partially supported by grant MTM2009-07848 (Spain). The second and fourth authors
are partially supported by the ANR project GEOMETRYA (France). Part of this work has
been realized during a research stay of the third author at the Université Paris Diderot (Paris
7) and Laboratory Jacques Louis Lions. The stay was supported by the program “Research in
Paris” offered by the Ville de Paris (Mairie de Paris). This author thanks the host institution
and Ville de Paris for its hospitality.
2. Notation and preliminaries
Let (Rn , d, L n ) denote the n-dimensional Euclidean space endowed with the Euclidean distance d(x, y) = kx − yk, the scalar product h·, ·i, and the Lebesgue measure L n . We denote
by B(x, r) (respectively, B(x, r)) the open (respectively, closed) ball of radius r > 0 and center
x ∈ Rn . If A is a nonempty subset of Rn , we denote by ]A its cardinality, by conv (A) its convex
hull and by diam A := sup {d(x, y) : x, y ∈ A} its diameter. We also denote by int(A), A and
∂A the interior, the closure and respectively, the boundary of the set A.
Let now K be a nonempty closed convex subset of Rn and u0 ∈ K. The normal cone NK (u0 )
is defined as follows:
NK (u0 ) = {v ∈ Rn : hv, u − u0 i ≤ 0, ∀u ∈ K}.
(2.1)
Notice that NK (u0 ) is always a closed convex cone. Notice further that u0 ∈ K is the projection
onto K of all elements of the form u0 + tv, where t ≥ 0 and v ∈ NK (u0 ).
The Hausdorff distance between two nonempty closed subsets K1 , K2 of Rn is given by the
formula
[
[
dH (K1 , K2 ) := inf ε > 0 : K2 ⊂
B(z, ε) and K1 ⊂
B(z, ε) .
z∈K1
z∈K2
A sequence of closed sets {Kj }j in Rn is said to converge with respect to the Hausdorff distance
to a closed set K ⊂ Rn if dH (Kj , K) → 0 as j → ∞.
RECTIFIABILITY OF SELF-CONTRACTED CURVES
IN THE EUCLIDEAN SPACE AND APPLICATIONS 5
Throughout the manuscript, I will denote a possibly unbounded interval of R. In this work, a
usual choice for the interval will be I = [0, T∞ ) where T∞ ∈ R ∪ {+∞}. A mapping γ : I → Rn
is referred in the sequel as a curve. Although the usual definition of a curve comes along with
continuity and injectivity requirements for the map γ, we do not make these prior assumptions
here. By the term continuous (respectively, absolutely continuous, Lipschitz, smooth) curve we
shall refer to the corresponding properties of the mapping γ : I → Rn . A curve γ is said to be
bounded if its image, denoted by Γ = γ(I), is a bounded set of Rn .
The length of a curve γ : I → Rn is defined as
n m−1
o
X
`(γ) := sup
d(γ(ti ), γ(ti+1 )) ,
(2.2)
i=0
where the supremum is taken over all finite increasing sequences t0 < t1 < · · · < tm that lie in
the interval I. Notice that `(γ) corresponds to the total variation of the function γ : I → Rn .
Let us mention for completeness that the length `(γ) of a continuous injective curve γ is equal
to the unidimensional Hausdorff measure H1 (Γ) of its image, see e.g. [3, Th.2.6.2], but it is
in general greater for noncontinuous curves. In particular we emphasis that for a piecewise
continuous curve, the quantity `(γ) is strictly greater than the sum of the lengths of each pieces
but we still call `(γ) the length of γ. A curve is called rectifiable, if it has locally bounded length.
Let us finally define the width of a (nonempty) convex subset K of Rn at the direction u ∈ Sn−1
as being the length of its orthogonal projection Pu (K) on the 1–dimensional space Ru generated
by u. The following definition will play a key role in our main result.
Definition 2.1 (Mean width). The mean width of a nonempty convex set K ⊂ Rn is given by
the formula
Z
1
L 1 (Pu (K))du,
W (K) =
σn Sn−1
where du denotes the standard volume form on Sn−1 , and
Z
nπ n/2
σn =
du =
.
Γ( n2 + 1)
Sn−1
2.1. Self-expanded curves. Let us now recall from [13] the definition and the basic properties
of a favorable class of Lipschitz curves, which has been studied thereby without a specific name.
In the sequel we call these curves self-expanded.
Definition 2.2 (Self-expanded curve). A Lipschitz curve γ : I → Rn is called
self-expanded
curve if for every t ∈ I such that γ 0 (t) exists, we have that γ 0 (t), γ(t) − γ(u) ≥ 0 for all u ∈ I
such that u ≤ t.
In [13, 3.IX.] the following result has been established concerning self-expanded curves.
Theorem 2.3 ([13, 3.IX.]). Let γ : I → Rn be a self-expanded curve of finite length. Then there
exists a constant C > 0 depending only on the dimension n such that
`(γ) ≤ C diam K,
(2.3)
where K is any compact set containing Γ = γ(I).
Notice that, formally, the above result requires the curve to have finite length. Nevertheless,
the following limiting argument allows to obtain a more general conclusion for bounded selfexpanded curves.
6
A. DANIILIDIS, G. DAVID, E. DURAND-CARTAGENA, A. LEMENANT
Corollary 2.4 (Bounded self-expanded curves have finite length). Every bounded self-expanded
curve γ : I → Rn has finite length and (2.3) holds.
Proof. Since self-expanded curves are rectifiable by definition, and because reparameterizing
γ does not change the statement, we may assume that γ is parameterized by its arc-length on
I = [0, `(γ)). Notice though that in principle `(γ) might be infinite. Our aim is precisely to
show that this is not the case. Indeed, let K be a compact set containing γ(I) and let {Ln }n be
an increasing sequence of real numbers converging to `(γ) ∈ R ∪ {+∞}. Applying Theorem 2.3
for the curve γn : [0, Ln ] → Rn (restriction of γ to [0, Ln ]), we obtain
Ln = `(γn ) ≤ C diam K,
for all n ≥ 1.
This shows that `(γ) = limn→+∞ Ln is bounded and satisfies the same estimate.
2.2. Self-contracted versus self-expanded curves. The aim of this section is to prove that
Lipschitz continuous self-contracted curves and self-expanded curves give rise to the same images. Moreover, each of these curves can be obtained from the other upon reparameterization
(inverting orientation). As a byproduct, bounded Lipschitz self-contracted curves in Rn have
finite length.
Let us recall the definition of a self-contracted curve (see [6, Definition 1.2]).
Definition 2.5 (Self-contracted curve). A curve γ : I → Rn is called self-contracted, if for every
t1 ≤ t2 ≤ t3 in I we have
d(γ(t1 ), γ(t3 )) ≥ d(γ(t2 ), γ(t3 )).
(2.4)
In other words, the function t 7→ d(γ(t), γ(t3 )) is nonincreasing on I ∩ (−∞, t3 ].
Remark 2.6. (i) As we already said before, the definition of self-contracted curve can be given
in any metric space and does not require any regularity of the curve, such as continuity or
differentiability. Notice moreover that if γ(t1 ) = γ(t3 ) in (2.4) above, then γ(t) = γ(t1 ) for
t1 ≤ t ≤ t2 ; thus if γ is not locally stationary, then it is injective.
(ii) It has been proved in [6, Proposition 2.2] that if γ is self-contracted and bounded, and
I = [0, T∞ ) with T∞ ∈ R+ ∪ {+∞}, then γ converges to some point γ∞ ∈ Rn as t → T∞ .
(Notice that this conclusion follows also from our main result Theorem 3.3.) Consequently the
curve γ can be extended to I¯ = [0, T∞ ]. In particular, if γ is continuous, then denoting by
Γ = γ(I) the image of γ, it follows that the set
¯
Γ̄ = γ(I) ∪ {γ∞ } = γ(I)
is a compact subset of Rn .
Lemma 2.7 (Property of a differentiability point). Let γ : I → Rn be a self-contracted curve
and let t be a point of differentiability of γ. Then
hγ 0 (t), γ(u) − γ(t)i ≥ 0 for all u ∈ I such that u > t.
Proof. Assume that γ is differentiable at t ∈ I, and write γ(t + s) = γ(t) + sγ 0 (t) + o(s), with
lims→0 s−1 o(s) = 0. Let u ∈ I be such that u > t, take s such that 0 < s < u − t and apply
(2.4) for t1 = t, t2 = t + s and t3 = u. We deduce that
kγ(t) − γ(u)k ≥ kγ(t + s) − γ(u)k.
RECTIFIABILITY OF SELF-CONTRACTED CURVES
IN THE EUCLIDEAN SPACE AND APPLICATIONS 7
Since γ(t + s) − γ(u) = γ(t) − γ(u) + sγ 0 (t) + o(s), substituting this in the above inequality and
squaring yields
0 ≥ kγ(t + s) − γ(u)k2 − kγ(t) − γ(u)k2
= 2hsγ 0 (t) + o(s), γ(t) − γ(u)i + ksγ 0 (s) + o(s)k2
= 2shγ 0 (t), γ(t) − γ(u)i + o(s).
Dividing by s, and taking the limit as s tends to 0+ we get the desired result.
Given a curve γ : I 7→ Rn , we denote by I − = −I = {−t ; t ∈ I} the opposite interval, and
define the reverse parametrization γ − : I − 7→ Rn of γ by γ − (t) = γ(−t) for t ∈ I − .
Lemma 2.8 (Lipschitz self-contracted versus self-expanded curves). Let γ : I → Rn be a
Lipschitz curve. Then γ is self-contracted if and only if γ − is a self-expanded curve.
Proof. If γ : I → Rn is Lipschitz and self-contracted, then Lemma 2.7 applies, yielding directly
that γ − is a self-expanded curve. Conversely, suppose that γ − is a self-expanded curve. This
means that γ is Lipschitz and (after reversing the orientation) that
hγ 0 (t), γ(u) − γ(t)i ≥ 0 for u ∈ I such that u > t
(2.5)
whenever γ 0 (t) exists and is different from 0.
Fix now any t3 ∈ I and define the function
1
f (t) = kγ(t) − γ(t3 )k2 , for all t ∈ I, t ≤ t3 .
2
By Rademacher’s theorem the Lipschitz continuous functions γ is differentiable L 1 -almost everywhere, and so f is too, and f 0 (t) = hγ 0 (t), γ(t) − γ(t3 )i for almost all t ∈ I ∩ (−∞, t3 ).
If t < t3 and γ 0 (t) 6= 0 then (2.5) above (for u = t3 ) yields that f 0 (t) ≤ 0. Otherwise,
= 0. It follows that the Lipschitz function f satisfies f 0 (t) ≤ 0 L 1 -almost everywhere,
thus it is nondecreasing. This establishes (2.4). Since t3 has been chosen arbitrarily, the proof
is complete.
f 0 (t)
3. Rectifiability of self-contracted curves
In [6, Theorem 1.3] it has been established that bounded self-contracted continuous planar
curves γ : [0, +∞) → R2 have finite length. In this section we improve this result by dropping
the continuity assumption, and we extend it to any dimension. Precisely, we establish that the
length of any self-contracted curve γ : I → Rn lying inside a compact set is bounded by a
quantity depending only on the dimension of the space and the diameter of the compact set, see
the forthcoming Theorem 3.3.
3.1. Proof of the main result. The proof makes use of the following technical facts.
Lemma 3.1 (Saturating the sphere). Let Σ ⊆ Sn−1 be such that hx, yi ≤ 1/2 for all x, y ∈
Σ, x 6= y. Then Σ is finite and ]Σ ≤ 3n .
Proof. For any x, y ∈ Σ, x 6= y, we have kx − yk2 = 2 − 2hx, yi ≥ 1. Therefore, the open balls
{B(x, 1/2)}x∈Σ are disjoint and they are all contained in the ball B(0, 23 ). Set ωn = L n (B(0, 1))
(the measure of the unit ball); then
n
n
[
3
1
n
n
=L
B(x, 1/2) ≤ L B(0, 3/2) = ωn
(]Σ) ωn
2
2
x∈Σ
8
A. DANIILIDIS, G. DAVID, E. DURAND-CARTAGENA, A. LEMENANT
and so ]Σ ≤ 3n .
Lemma 3.2 (Hemisphere lemma). Let Σ ⊂ Sn−1 be a set satisfying
n+1
1
hx, yi ≥ −
for all x, y ∈ Σ.
3
(3.1)
Then there exists ζ ∈ Sn−1 such that
2n+1
1
for all x ∈ Σ.
hζ, xi ≥
3
Proof. Let {xi }i∈I be a family of points in Σ, maximal with respect to the property that
1
for all i 6= j in I .
(3.2)
2
(Notice that such a family can easily be constructed by induction.) Applying Lemma 3.1 we
deduce that m := ]I ≤ 3n . Set
X
v :=
xi
hxi , xj i ≤
i∈I
and let y ∈ Σ be an arbitrary point. If y does not belong to the family {xi }i∈I then by maximality
of the latter, there exists some i0 such that hxi0 , yi > 21 . If on the contrary, y = xi for some i,
then we take i0 = i. In view of (3.1) we obtain
n+1
X
1
1
,
(3.3)
hv, yi = hxi0 , yi +
hxi , yi ≥ − (m − 1)
2
3
i6=i0
whence, recalling that m ≤ 3n , we get hy, vi ≥ 3−(n+1) . This shows in particular that v 6= 0.
Let us now compute kvk. We write
X 2 X
2
kvk = xi =
kxi k2 + E = m + E
(3.4)
i∈I
i∈I
where
E=
XX
hxi , xj i.
i∈I j6=i
Using (3.2) we deduce that |E| ≤ m(m − 1)/2. Therefore (3.4) yields
m
3n n
m(m − 1)
= (m + 1) ≤
(3 + 1) ≤ 32n .
2
2
2
So dividing by kvk in (3.3) and setting
v
ζ :=
∈ Sn−1
kvk
kvk2 ≤ m +
we obtain
hζ, yi =
1
hv, yi ≥
kvk
n n+1 2n+1
1
1
1
=
.
3
3
3
Since y is arbitrary in Σ, the proof is complete.
We are now ready to prove the main result of this section.
RECTIFIABILITY OF SELF-CONTRACTED CURVES
IN THE EUCLIDEAN SPACE AND APPLICATIONS 9
Theorem 3.3. Let γ : I → Rn be a self-contracted curve. Then there exists a constant Cn
(depending only on the dimension n) such that
`(γ) ≤ Cn W (K),
(3.5)
where K is the closed convex hull of γ(I). In particular, bounded self-contracted curves have
finite length.
Proof. The result holds vacuously for unbounded curves (both left-hand and right-hand side
of (3.5) are equal to +∞). Therefore, we focus our attention on bounded self-contracted curves
and assume that K is compact. We may also clearly assume that n ≥ 2 (the result is trivial in
the one-dimensional case).
In the sequel, we denote by Γ = γ(I) the image of such curve. The set Γ inherits from I a
total order as follows: for x, y ∈ Γ we say that “x is before y ” and denote x y, if there exist
t1 , t2 ∈ I, t1 ≤ t2 and γ(t1 ) = x, γ(t2 ) = y. If x y and x 6= y, then the intervals γ −1 (x) and
γ −1 (y) do not meet, and for any t1 ∈ γ −1 (x), t2 ∈ γ −1 (y) we have t1 < t2 . In this case we say
that “x is strictly before y” and we denote x ≺ y. For x ∈ Γ we set
Γ(x) := {y ∈ Γ : x y}
(the piece of curve after x) and denote by Ω(x) the closed convex hull of Γ(x).
Claim 1. To establish (3.5) it suffices to find a positive constant ε = ε(n), depending only on
the dimension n, such that for any two points x, x0 ∈ Γ with x0 x it holds
W (Ω(x)) + εkx − x0 k ≤ W (Ω(x0 )).
(3.6)
Proof of Claim 1. Let us see how we can deduce Theorem 3.3 from the above. To this end,
let t0 < t1 . . . < tm be any increasing sequence in I, and set xi = γ(ti ). If (3.6) holds, then
m−1
X
i=0
kγ(ti+1 ) − γ(ti )k =
m−1
X
i=0
kxi+1 − xi k ≤
m−1
1X
W (C(xi ) − W (C(xi+1 ))
ε
i=0
1
1
1
= (W (C(x0 )) − W (C(xm ))) ≤ W (C(x0 )) ≤ W (K),
ε
ε
ε
since the mean width W (H) is a nondecreasing function of H (the variable H is ordered via the
set inclusion). Taking the supremum over all choices of t0 < t1 . . . < tm in I we obtain (3.5) for
Cn = 1/ε.
N
Therefore, the theorem will be proved, if we show that (3.6) holds for some constant ε > 0
which depends only on the dimension. Before we proceed, we introduce some extra notation.
Given x, x0 in Γ with x0 ≺ x we set
1
2
x0 − x
x0 := x + x0 and v0 := 0
.
3
3
kx − xk
(3.7)
For the sake of drawing pictures, the reader is invited to think that v0 = e1 (the first vector of
the canonical basis of Rn ), see Figure 1. Let us also set
y − x0
ξ0 (y) =
∈ Sn−1 , for any y ∈ Γ(x).
(3.8)
ky − x0 k
Clearly, x0 , v0 and ξ0 (y) depend on the points x, x0 , while the desired constant ε does not.
To determine this constant, we shall again transform the problem into another one (see the
forthcoming Claim 2).
10
A. DANIILIDIS, G. DAVID, E. DURAND-CARTAGENA, A. LEMENANT
Figure 1. Controlling the tail of a self-contracted curve.
Claim 2. Let us assume that there exists a constant 0 < δ < 2−4 , depending only on the
dimension n, such that for all x, x0 in Γ with x0 ≺ x (and for x0 , v0 defined by (3.7)), there exists
v̄ ∈ Sn−1 ∩ B(v0 , δ) such that
v, ξ0 (y) ≤ −δ 2 for all y ∈ Γ(x).
(3.9)
Then (3.6) holds true (and consequently (3.5) follows).
Proof of Claim 2. Assume that such a constant δ and a vector v̄ exist, so that (3.9) holds.
Set
V = v ∈ Sn−1 ; kv − vk ≤ δ 2 .
Combining with (3.8) and (3.9) we get
hv, y − x0 i ≤ 0,
for all v ∈ V and y ∈ Γ(x).
(3.10)
Let us first explain intuitively why the above yields (3.6). Indeed, compared with Ω(x), Ω(x0 ) has
an extra piece coming from the segment [x0 , x0 ]. This piece is protruding in all directions v ∈ V
(which are relatively close to v0 ), while (3.10) bounds uniformly the orthogonal projections of
Ω(x) onto the lines Rv. Therefore, the extra contribution of the segment [x0 , x0 ] in Pv (Ω(x0 ))
becomes perceptible and can be quantified in terms of kx0 − x0 k, uniformly in V. Since the latter
set V has a positive measure, the estimation (3.6) follows.
To proceed, observe that Ω(x0 ) contains the convex hull of Ω(x) ∪ [x, x0 ], whence
Pv (Ω(x)) ⊂ Pv (Ω(x0 )).
In particular,
H1 (Pv (Ω(x))) ≤ H1 (Pv (Ω(x0 ))) for v ∈ Sn−1 ,
(3.11)
where Pv denotes the orthogonal projection onto the line Rv. Let us now equip the latter with
the obvious order (stemming from the identification Rv ∼
= R) and let us identify Pv (x0 ) with
the zero element 0. Then (3.10) says that for all directions v in V we have
sup Pv (Ω(x)) ≤ 0 < Pv (x0 ) ≤ sup Pv (Ω(x0 )).
Notice that
1
kx0 − x0 k = kx − x0 k,
3
RECTIFIABILITY OF SELF-CONTRACTED CURVES
IN THE EUCLIDEAN SPACE AND APPLICATIONS 11
and that V ⊂ B(v0 , δ + δ 2 ). Thus for every v ∈ V we have
hv0 , vi ≥ hv0 , v0 i − kv0 kkv0 − vk ≥ 1 − δ − δ 2 ≥ 7/8.
This gives a lower bound for the length of the projected segment [x0 , x0 ] onto Rv, which coincides,
under the above identification, with Pv (x0 ). Thus
7 1
1
0
0
Pv (x ) ≥
kx − x k > kx − x0 k.
8 3
4
This yields
1
(3.12)
H1 (Pv (Ω(x))) + kx − x0 k ≤ H1 (Pv (Ω(x0 ))) for v ∈ V .
4
Integrating (3.12) for v ∈ V and (3.11) for v ∈ Sn−1 \ V , and summing up the resulting
inequalities we obtain (3.6) with
Z
ε = (4σn )−1
du.
V
Notice that this bound only depends on δ, so the claim follows.
N
Consequently, our next goal is to determine δ > 0 so that the assertion of Claim 2 holds. The
exact value of the parameter δ is eventually given in (3.25) and depends only on the dimension.
In particular, it works for any self-contracted curve and any choice of points x, x0 ∈ Γ. For the
remaining part of the proof, it is possible to replace δ by its precise value. Nevertheless, we
prefer not to do so, in order to illustrate how this value is obtained. In the sequel, the only prior
requirement is the bound δ ≤ 2−4 .
Fix any x, x0 in Γ with x0 ≺ x and recall the definition of x0 , v0 in (3.7) and ξ0 (y), for y ∈ Γ(x)
in (3.8). Based on this, we consider the orthogonal decomposition
Rn ≈ Rv0 ⊕ U
(3.13)
where U = (v0 )⊥ is the orthogonal hyperplane to v0 . The vector v of the assertion of Claim 2
will be taken of the form
v0 − δζ
v=
,
(3.14)
kv0 − δζk
where ζ is a unit vector in U. This vector will be determined later on, as an application of
Lemma 3.2; notice however that for any ζ ∈ Sn−1 ∩ U we get v ∈ Sn−1 ∩ B(v0 , δ), as needed.
We set
Γ0 = y ∈ Γ(x) ; hv0 , ξ0 (y)i ≤ −2δ .
(3.15)
Notice that for every y ∈ Γ0 and any v of the form (3.14) we have
hv̄, ξ0 (y)i ≤ −2δ + kv0 − v̄k ≤ −δ ≤ −δ 2 .
Thus (3.9) is satisfied for all y in Γ0 .
It remains to choose ζ (and adjust the value of δ) so that (3.9) would also hold for y ∈ Γ(x)\Γ0 .
This will be done in five steps. In the sequel we shall make use of the decomposition (3.13) of
vectors ξ0 (y), y ∈ Γ(x), namely:
ξ0 (y) = hv0 , ξ0 (y)i v0 + ξ0U (y),
where ξ0U (y) is the orthogonal projection of ξ0 (y) in U.
(3.16)
12
A. DANIILIDIS, G. DAVID, E. DURAND-CARTAGENA, A. LEMENANT
Step 1. We establish that for all y ∈ Γ(x) \ {x} it holds
hv0 , ξ0 (y)i ≤ −
kx − x0 k
< 0.
6ky − x0 k
(3.17)
Indeed, since x0 ≺ x ≺ y, we deduce from self-contractedness that ky − x0 k ≥ ky − xk, that is,
y lies at the same half-space as x defined by the mediatrix hyperplane of the segment [x, x0 ].
Hence, denoting by Pv0 the orthogonal projection of Rn on Rv0 (which we brutally identify to
R to write inequalities) we observe that
1
1
1
Pv0 (y) ≤ Pv0 (x) + Pv0 (x0 ) = Pv0 (x) + kx − x0 k,
2
2
2
and consequently
1
1
hv0 , y − x0 i = Pv0 (y − x0 ) ≤ Pv0 (x) + kx − x0 k − Pv0 (x0 ) = − kx − x0 k,
2
6
thus, dividing by ky − x0 k, (3.17) follows.
Step 2. We establish that for all y ∈ Γ(x) \ Γ0 , we have:
1
ky − x0 k >
kx − x0 k ;
12δ
|hv0 , ξ0 (y)i| = kξ0 (y) − ξ0U (y)k ≤ 2δ ;
and
p
1 − 4δ 2 ≤ kξ0U (y)k ≤ 1 .
(3.18)
(3.19)
(3.20)
Indeed, since y ∈ Γ(x) \ Γ0 , (3.18) follows easily by combining (3.15) with (3.17). The same
formulas yield that hv0 , ξ0 (y)i ∈ (−2δ, 0), whence |hv0 , ξ0 (y)i| ≤ 2δ. In view of (3.16), both (3.19)
and (3.20) follow directly. This ends the proof of Step 2.
Before we proceed, let us make the following observation, which motivates Step 4 (compare
forthcoming inequalities (3.22) and (3.24)). We set
y−x
∈ Sn−1 , for all y ∈ Γ(x) \ {x}.
(3.21)
ξ(y) =
ky − xk
It is easily seen that
hξ(y), ξ(z)i ≥ 0, for all y, z ∈ Γ(x) \ {x}.
(3.22)
Indeed assuming x y ≺ z (the case y = z is trivial), the definition of self-contractedness yields
kz − xk ≥ kz − yk.
Writing z − y = (z − x) − (y − x) and squaring the above inequality, we obtain
kz − xk2 ≥ kz − xk2 + kx − yk2 − 2hz − x, y − xi,
which yields hz − x, y − xi ≥ 0. Dividing by norms gives (3.22).
Our next objective is to show that vectors ξ0U (y), y ∈ Γ(x) \ Γ0 satisfy a relaxed inequality of
type (3.22). We need the following intermediate step.
Step 3. We show that for every y ∈ Γ(x) \ Γ0 , we have:
kξ(y) − ξ0 (y)k ≤ 32 δ.
(3.23)
RECTIFIABILITY OF SELF-CONTRACTED CURVES
IN THE EUCLIDEAN SPACE AND APPLICATIONS 13
Indeed
y−x
y − x0 kξ(y) − ξ0 (y)k ≤ −
+
ky − xk ky − xk
kx − x0 k
1
=
+ |
−
ky − xk
ky − xk
2kx − x0 k
4 kx − x0 k
≤
=
.
ky − xk
3 ky − xk
y−x
y − x0 0
−
ky − xk ky − x0 k
1
| ky − x0 k
ky − x0 k
On the other hand, using (3.18) we get:
1
2
− )kx − x0 k =
12δ 3
Combining the above inequalities, we deduce
16δ
kξ(y) − ξ0 (y)k ≤
,
1 − 8δ
thus (3.23) follows thanks to the prior bound δ ≤ 2−4 .
ky − xk ≥ ky − x0 k − kx0 − xk ≥ (
1 − 8δ
16δ
4kx − x0 k
3
.
Step 4. We now establish that for every y, z ∈ Γ(x) \ Γ0 , we have:
hξ0U (y), ξ0U (z)i ≥ −65 δ.
(3.24)
Let us first calculate the scalar product of the unit vectors ξ0 (y) and ξ0 (z).
hξ0 (y), ξ0 (z)i = hξ(y), ξ0 (z)i + hξ0 (y) − ξ(y), ξ0 (z)i
= hξ(y), ξ(z)i + hξ(y), ξ0 (z) − ξ(z)i + hξ0 (y) − ξ(y), ξ0 (z)i
≥ 0 − kξ0 (z) − ξ(z)k − kξ0 (y) − ξ(y)k ≥ −64 δ ,
where (3.22) and (3.23) have been used for the last estimation.
Using now the above estimation, the orthogonal decomposition (3.16) recalling (3.19) and the
prior bound δ ≤ 2−4 we obtain successively:
hξ0U (y), ξ0U (z)i = hξ0 (y), ξ0 (z)i − hv0 , ξ0 (y)ihv0 , ξ0 (z)i
≥ −64δ − 4δ 2 ≥ −65δ.
Step 5. We shall now be interested in the set
Σ = ξˆ0U (y) ; y ∈ Γ(x) \ Γ0 ⊂ U ∩ Sn−1 ,
where ξˆ0U (y) = ξ0U (y)/kξ0U (y)k (notice that ξ0U (y) 6= ∅ from (3.20)). In view of (3.24) and the
estimation (3.20), we obtain that for all y, z ∈ Γ(x) \ Γ0
−65δ
hξˆ0U (y), ξˆ0U (z)i ≥
≥ −81δ.
1 − 4δ 2
We are now ready to give a precise value for δ, namely,
1 3n
δ=
.
(3.25)
3
Recalling that n ≥ 2, under this choice the scalar product between two elements of Σ is bounded
from below by the quantity
1 3n
1 n
−81 δ ≥ −81
≥−
.
3
3
14
A. DANIILIDIS, G. DAVID, E. DURAND-CARTAGENA, A. LEMENANT
Therefore, the subset Σ of the unit sphere of U ' Rn−1 satisfies the assumptions of Lemma 3.2.
We deduce that there exists a unit vector ζ ∈ U ∩ Sn−1 such that
1 2n−1
hζ, ξˆ0U (y)i ≥
, for all y ∈ Γ(x) \ Γ0 .
(3.26)
3
Taking δ and ζ as above in (3.14), we obtain v ∈ Sn−1 ∩ B(v0 , δ). Let us now verify that (3.9)
holds. As already mentioned (see comments before Step 1), it remains to consider the case
y ∈ Γ(x) \ Γ0 . Notice that in view of (3.17)
hv0 − δζ, ξ0 (y)i = hv0 , ξ0 (y)i − δ hζ, ξ0 (y)i ≤ −δ hζ, ξ0 (y)i,
while in view of (3.13), (3.20) and the orthogonality between v0 and ζ we get
p
1 2n−1
1 2n
hζ, ξ0 (y)i = hζ, ξ0U (y)i = kξ0U (y)k hζ, ξˆ0U (y)i ≥ 1 − 4δ 2
≥
.
3
3
√
Assembling the above, and using the fact that kv0 − δζk = 1 + δ 2 > 1 we obtain
1 2n
1 3n
p
hv̄, ξ0 (y)i ≤ −δ 1 + δ 2
≤ −δ
= −δ 2 .
3
3
This shows that the assumption made in Claim 2 is always fulfilled. The proof is complete.
3.2. Arc-length reparameterization. Having in hand that the length of a bounded selfcontracted curve γ : I → Rn is finite (c.f. Theorem 3.3), the curve γ may be reparameterized by
its arc-length. This reparameterization is particularly interesting when the curve γ is continuous.
In this case, we shall see that the arc-length parametrization is Lipschitz continuous and that
the image Γ = γ(I) of the curve can also be realized by a self-expanded curve.
Set I(t) = I ∩ (−∞, t], for t ∈ I, and consider the length function
L : I → [0, `(γ)]
L(t) = `(γ|I(t) ).
Thus L(t) is the length of the truncated curve γ|I(t) and consequently the function t 7→ L(t)
is nondecreasing, but not necessarily injective —it is locally constant whenever t 7→ γ(t) is
locally stationary, see also Remark 2.6 (i). Let us assume that γ is bounded and continuous.
Then, in view of Remark 2.6 (ii), the curve can be extended continuously to I¯ and the set
¯ is a compact connected arc. Moreover, t 7→ L(t) is continuous and
Γ̄ = γ(I) ∪ {γ∞ } = γ(I)
¯ = [0, `(γ)]. It follows easily that L−1 (s) is a (possibly trivial) interval [t1 , t2 ]. Consequently,
L(I)
the function
γ̃ : [0, `(γ)] → Rn
(3.27)
γ̃(s) = γ(t), for any t ∈ L−1 (s)
is well-defined.
Proposition 3.4 (Hidden regularity of continuous self-contracted curves). Let γ : I → Rn
be a bounded continuous self-contracted curve. Then (3.27) defines a Lipschitz self-contracted
curve with the same image Γ = γ(I) = γ̃([0, `(γ))). In particular, Γ is also the image of some
self-expanded curve.
Proof. Let s, s0 ∈ [0, `(γ)] be given. Let t, t0 ∈ I be such that s = L(t) and s0 = L(t0 ). Assume
that s < s0 (thus a fortiori t < t0 ). We claim that
L(t0 ) ≥ L(t) + d(γ(t), γ(t0 )).
(3.28)
RECTIFIABILITY OF SELF-CONTRACTED CURVES
IN THE EUCLIDEAN SPACE AND APPLICATIONS 15
Indeed, recalling (2.2), for any ε > 0, there exists a finite sequence t0 < . . . < tm in I(t),
satisfying
m−1
X
Σ1 =:
d(γ(ti ), γ(ti+1 )) ≥ L(t) − ε.
i=0
Since adding an extra point can only make the sum larger, we may assume that tm = t. Then
add the point tm+1 = t0 and notice that t0 < . . . < tm < tm+1 is a sequence in I(t0 ), thus
m
X
Σ2 =:
d(γ(ti ), γ(ti+1 )) = Σ1 + d(γ(t), γ(t0 )) ≤ L(t0 ).
i=0
We deduce that
L(t) − ε ≤ Σ1 = Σ2 − d(γ(t), γ(t0 )) ≤ L(t0 ) − d(γ(t), γ(t0 )).
Taking the limit as ε → 0 and get (3.28). It follows that
kγ̃(s0 ) − γ̃(s)k = kγ(t0 ) − γ(t)k ≤ L(t0 ) − L(t) = s0 − s,
which proves that γ̃ is 1-Lipschitz on [0, `(γ)].
It follows easily that γ̃ is self-contracted, while the fact that γ̃([0, `(γ))) = γ(I) is straightforward. The last assertion follows from Lemma 2.8.
Remark 3.5 (Improving the length bound). Although the constants C > 0 in (2.3) (for selfexpanded curves) and in (3.5) (for bounded self-contracted curves) stem from the same method,
the latter is based on nonsmooth arguments encompassing nonregular curves and therefore it
is not optimized. As a consequence of Proposition 3.4, given a continuous self-contracted curve
we can use (2.3) to improve a posteriori the constant that bounds the length given in (3.5).
Notwithstanding, this trick cannot be applied to any bounded self-contracted curve: in the
discontinuous case, one can still consider the length mapping L(t), and use it to construct a
larger curve (whose image contains Γ) with a Lipschitz parametrization, but which is possibly
not self-contracted.
4. Applications
4.1. Orbits of convex foliations. As mentioned in the introduction, self-contracted curves
are naturally defined in a metric space without prior regularity assumptions. Nonetheless, this
notion has been conceived to capture the behaviour of orbits of gradient dynamical systems
for (smooth) quasiconvex potentials in Euclidean spaces (1.1), or more generally, orbits of the
subgradient semi-flow defined by a nonsmooth convex function, see (1.2). In both cases, orbits
are at least, absolutely continuous. In this section we broaden the above framework by introducing a new concept of generalized solution to a nonsmooth convex foliation, admitting (merely)
continuous curves as generalized orbits.
4.1.1. Nonsmooth convex foliations. Let us first give the definition of a convex foliation.
Definition 4.1 (Convex foliation). A collection {Cr }r≥0 of nonempty convex compact subsets
of Rn is a (global) convex foliation of Rn if
r1 < r2 =⇒ Cr1 ⊂ int Cr2
(4.1)
[
(4.2)
and
r≥0
∂Cr = Rn int C0 .
16
A. DANIILIDIS, G. DAVID, E. DURAND-CARTAGENA, A. LEMENANT
In view of (4.1), relation (4.2) is equivalent to the fact that for every x ∈ Rn \ int C0 , there
exists a unique r ∈ [0, ∞) such that x ∈ ∂Cr . The above definition is thus equivalent to [6,
Definition (6.5)].
Example 4.2 (Foliation given by a function). (i) The sublevel sets of a proper (coercive) convex
function f : Rn → R provide a typical example of a convex foliation. Indeed, if m = min f , we
set C0 = f −1 (m) and Cr = [f ≤ m + r]. Then (4.2) follows from the fact that f cannot be
locally constant outside of C0 .
(ii) The sublevel sets of a (coercive) quasiconvex function might fail to satisfy (4.2), unless
int [f ≤ λ] = [f < λ] for all λ > min f. This condition is automatically satisfied if the quasiconvex
function is smooth and every critical point is a global minimizer. A more general condition is
to assume that f is semistrictly quasiconvex, see for instance [5] for the relevant definition and
characterizations.
4.1.2. Generalized solutions and self-contractedness. We consider a possibly unbounded interval
I = [0, T∞ ) of R, where T∞ ∈ R+ ∪ {+∞} and a continuous injective curve γ : I → Rn . For
every τ ∈ I we define the set of all possible limits of backward secants at γ(τ ) as follows:
γ(tk ) − γ(τ )
−
n−1
sec (τ ) := q ∈ S
: q = lim
,
(4.3)
tk %τ − kγ(tk ) − γ(τ )k
where the notation {tk }k % τ − indicates that {tk }k → τ and tk < τ for all k. The set
sec+ (τ ) is defined analogously using decreasing sequences {tk }k & τ + . The compactness of
Sn−1 guarantees that both sec− (τ ) and sec+ (τ ) are nonempty. The following lemma relates
sec− (τ ) and sec+ (τ ) with the derivative of γ, in case the latter exists and does not vanish.
Lemma 4.3 (Secants versus derivative). Assume γ : I → Rn is differentiable at τ ∈ I and
γ 0 (τ ) 6= 0. Then
0
γ (τ )
γ 0 (τ )
and sec+ (τ ) =
.
sec− (τ ) = − 0
kγ (τ )k
kγ 0 (τ )k
Proof. Let q ∈ sec− (τ ) and let {tk }k % τ − be a sequence realizing this limit. Then writing
γ(tk ) − γ(τ )
γ(tk ) − γ(τ )
−|τ − tk |
=
kγ(tk ) − γ(τ )k
tk − τ
kγ(tk ) − γ(τ )k
and passing to the limit as {tk }k % τ − we obtain the result. The second assertion follows
similarly.
Assuming γ(t) ∈
/ int C0 we denote by r(t) the unique positive number such that
γ(t) ∈ ∂Cr(t) .
(4.4)
We are now ready to give the definition of generalized orbit to a convex foliation.
Definition 4.4 (Convex foliation orbits). A continuous curve γ : I → Rn is called generalized
solution (orbit) of the convex foliation {Cr }r≥0 if γ(I) ∩ int C0 = ∅ and under the notation (4.3)
and (4.4) we have:
(i) the function t 7→ r(t) is decreasing ;
(ii) sec− (t) ⊂ NCr(t) (γ(t)) for all t ∈ I.
RECTIFIABILITY OF SELF-CONTRACTED CURVES
IN THE EUCLIDEAN SPACE AND APPLICATIONS 17
It follows directly from (i) that γ is injective, and then (ii) makes sense. Also, every orbit of
a convex foliation is an injective mapping with bounded image (Γ = γ(I) is contained in the
compact set Cr(0) ). Recall the definition of the normal cone NCr(t) in (2.1). If ∂Cr(t) is a smooth
manifold (or more generally, if dim NCr(t) (γ(t)) = 1) we have sec− (t) = {νr(t) } where νr(t) is the
unit normal of ∂Cr(t) at γ(t).
In the sequel, we shall need the following lemma, which proof is an easy exercise and will be
omitted.
Lemma 4.5 (Criterium for decrease). Let ϕ : [0, T ] → R be a continuous function, where T ≥ 0.
Assume that for every 0 < τ ≤ T, there exists δ > 0 such that for all t ∈ (τ − δ, τ ) ∩ [0, T ] we
have
ϕ(t) > ϕ(τ ).
(4.5)
Then ϕ is decreasing.
We now prove that the generalized solutions of a convex foliation are self-contracted curves.
Notice that in view of Theorem 3.3 this entails that these curves are rectifiable and have a finite
length.
Theorem 4.6 (Self-contractedness of generalized orbits). Every convex foliation orbit is a (continuous) self-contracted curve.
Proof. Let γ : I → Rn be a continuous injective curve which is a generalized solution of the
convex foliation {Cr }r≥0 in the sense of Definition 4.4. Since γ(I) ⊂ Cr(0) , the curve is obviously
bounded. Fix τ ∈ I and ε > 0 such that τ + ε ∈ I. We set
Γε := {γ(t) : t ≥ τ + ε, t ∈ I}.
Notice that the compact set Γε is contained in int Cr(τ +ε0 ) for all 0 ≤ ε0 < ε.
Claim 1. There exists a > 0, such that for all v ∈ NCr(τ ) (γ(τ )) and all x ∈ Γε we have
h v,
x − γ(τ )
i < −a < 0 .
kx − γ(τ )k
(4.6)
Proof of Claim 1. Assume, towards a contradiction, that (4.6) fails. Since the sets Γ and
NCr(τ ) (γ(τ )) ∩ Sn−1 are compact, we easily deduce that for some unit normal v ∈ NCr(τ ) (γ(τ ))
and x ∈ Γ we have hv, x − γ(τ )i ≥ 0. Since x ∈ int Cr(τ ) , we deduce that for some y ∈ Cr(τ ) we
have hv, y − γ(τ )i > 0. This contradicts the fact that v ∈ NCr(τ ) (γ(τ )). Thus (4.6) holds.
N
Claim 2. For every ε > 0 there exists δ > 0 such that for all t− , t+ ∈ I such that t− ∈ [τ −δ, τ )
and t+ ≥ τ + ε we have
hγ(t− ) − γ(τ ), γ(t+ ) − γ(τ )i < 0 .
(4.7)
Proof of Claim 2. Fix ε > 0 and let a > 0 be given by Claim 1. Let N 1 = NCr(τ ) (γ(τ )) ∩ Sn−1
denote the (compact) set of unit normals at γ(τ ), let Uα := N 1 +B(0, α) be its α-enlargement, and
let Nα denote the closed convex cone generated by Uα . It follows from (4.6) that if α > 0 is chosen
small enough (depending on a), then for all w ∈ Nα , and all x ∈ Γ , we have hw, x − γ(τ )i < 0.
Since sec− (τ ) ⊂ N 1 ⊂ Ua , we deduce that for some δ > 0 we have
γ(t) − γ(τ )
∈ Uα ,
kγ(t) − γ(τ )k
This establishes (4.7).
for all t ∈ (τ − δ, τ ) ∩ I.
N
18
A. DANIILIDIS, G. DAVID, E. DURAND-CARTAGENA, A. LEMENANT
Notice that (4.7) yields
kγ(t− ) − γ(t+ )k > kγ(τ ) − γ(t+ )k.
(4.8)
We now prove that the curve γ is self-contracted. Indeed, fix 0 ≤ t1 < t2 < t3 in I and consider
the real-valued function
(
ϕ : [0, t2 ] → R
ϕ(t) = kγ(t) − γ(t3 )k
Applying Claim 2 for any 0 < τ ≤ t2 and for ε = t3 − t2 we deduce that for some δ1 > 0 and
all t ∈ (τ − δ1 , τ ) ∩ [0, t2 ] we have ϕ(t) > ϕ(τ ). The conclusion follows from Lemma 4.5. Since
t1 , t2 , t3 are arbitrarily chosen we deduce that γ is self-contracted.
As a consequence of Theorem 4.6 and [6, Proposition 2.2] the solution curve γ has finite
length. In particular, the curve converges as t → T∞ to some limit point
γ∞ := lim γ(t) ∈ ∂Cr(T∞ )
t→T∞
so that the mapping γ can be continuously extended to [0, T∞ ], by setting γ(T∞ ) := γ∞ .
Remark 4.7 (Generalized versus classical solutions). Let γ : I → Rn be a convex foliation
orbit and Γ = γ(I). Although γ is merely continuous, Theorem 4.6 guarantees its rectifiability.
Thus Proposition 3.4 applies, and the curve can be reparameterized by its arc-length to obtain
a Lipschitz curve with the same image. The new curve γ∗ : [0, `(γ)] → Rn satisfies kγ∗0 (s)k = 1
and
−γ∗0 (s) ∈ NC(r(s)) (γ∗ (s))
for almost all s ∈ [0, γ(`)]. It follows that every generalized orbit of a convex foliation gives rise
to a classical solution (Lipschitz orbit).
4.1.3. Smooth solutions and strong self-contractedness. Let γ : I → Rn be a continuous curve,
where I = [0, T∞ ), let T ∈ I and denote by Ω(T ) the convex hull of the tail of the curve, that is,
Ω(T ) := conv {γ(t) : t ≥ T }.
Let us give the following definitions.
Definition 4.8 (Strong self-contractedness). A continuous injective curve γ : I → Rn is called
• strictly self-contracted, if for t1 , t2 , t3 in I such that t1 < t2 < t3 we have
kγ(t1 ) − γ(t3 )k > kγ(t2 ) − γ(t3 )k.
• strongly self-contracted, if for all τ ∈ I we have
sec− (τ ) ⊂ int NΩ(τ ) (γ(τ )).
(4.9)
Of course every strictly self-contracted curve is self-contracted. A careful look at the proof of
Theorem 4.6 actually reveals that the orbits of the convex foliation are strictly self-contracted
curves. The following proposition relates the notions of strict and strong self-contractedness.
Proposition 4.9 (Strong versus strict self-contractedness). Every strongly self-contracted curve
is strictly self-contracted.
RECTIFIABILITY OF SELF-CONTRACTED CURVES
IN THE EUCLIDEAN SPACE AND APPLICATIONS 19
Proof. Fix t1 < t2 < t3 in I and apply (4.9) for τ = t2 . From the definition of sec− (τ ), see
(4.3), we deduce that for some δ > 0 and all t ∈ (t2 − δ, t2 ) we have
hγ(t) − γ(t2 ), γ(t3 ) − γ(t2 )i < 0.
Setting ϕ(t) = kγ(t) − γ(t3 )k, the above inequality guarantees that Lemma 4.5 applies. The
proof is complete.
Our next objective is to show that smooth orbits of a convex foliation are actually strongly
self-contracted curves. We shall need the following lemma.
Lemma 4.10 (Differentiability point of a convex foliation orbit). Let γ : I → Rn be a continuous
orbit of a convex foliation and assume that γ is differentiable at some τ > 0 with γ 0 (τ ) 6= 0.
Then
−γ 0 (τ ) ∈ int NΩ(τ ) (γ(τ )).
(4.10)
Proof. In order to simplify notation we assume γ(τ ) = 0. (There is no loss of generality in
doing this.) In view of Lemma 4.3 there exist ε > 0, such that for all t+ ∈ (τ, τ + ε) we have
h
γ(t+ )
1
γ 0 (τ )
,
i > .
0
+
kγ (τ )k kγ(t )k
2
(4.11)
Since the compact set Ω(τ + ε) is contained in int Cr(τ ) , arguing as in Claim 1 of the proof of
Theorem 4.6, we deduce that for some a > 0 and for all t+ ≥ τ + ε we have
h−
γ 0 (τ )
γ(t+ )
,
i < −a.
kγ 0 (τ )k kγ(t+ )k
(4.12)
Changing a into min(a, 1/2) if necessary, we deduce that (4.12) holds true for all t+ > τ . We
deduce that the tail of the curve for t ≥ τ is contained in the convex cone
o
n
γ 0 (τ )
, xi ≤ −akxk ,
x ∈ Rn : h 0
kγ (τ )k
hence so does Ω(τ ). This easily yields −γ 0 (τ ) ∈ int NΩ(τ ) (γ(τ )).
Corollary 4.11 (Strong self-contractedness of smooth orbits). Every C 1 convex foliation orbit
with no stationary point is a strongly self-contracted curve.
Proof. Parameterizing the curve by its arc-length parametrization we obtain a C 1 curve with
nonvanishing derivative and the same image Γ = γ(I). This new curve satisfies kγ̇(s)k = 1, for
every s ∈ [0, `(γ)] and it is also a convex foliation orbit, see Remark 4.7. The result follows by
combining Lemma 4.3 with Lemma 4.10.
4.2. Polygonal approximations of smooth strongly self-contracted curves. In this section we prove that every strongly self-contracted C 1 curve is a limit of self-contracted polygonal
curves (with respect to the Hausdorff distance). Let us recall the relevant definition.
Definition 4.12 (Polygonal approximation). Let γ : I → Rn be a continuous curve. A polygonal
m
S
line P =
[zk+1 , zk ] is called polygonal approximation of accuracy δ > 0 for the curve γ, if
k=0
{zk }k ⊂ γ(I),
zk+1 zk
and
dH (P, γ(I)) ≤ δ.
20
A. DANIILIDIS, G. DAVID, E. DURAND-CARTAGENA, A. LEMENANT
In the above we used the same notation x y which corresponds to the order on the curve as
in the beginning of the proof of Theorem 3.3. In the sequel, bounded continuous self-contracted
¯ so that Γ = γ(I)
¯ is a compact set.
curves γ : I → Rn are considered to be extended to I,
Proposition 4.13 (Self-contracted polygonal approximations). Let γ : [0, L] → Rn be a C 1
strongly self-contracted curve. Then for every δ > 0 the curve γ admits a self-contracted polygonal approximation of accuracy δ > 0.
Proof. Since the statement is independent of the parametrization, there is no loss of generality
in assuming that the curve is parameterized by arc-length. In particular, kγ 0 (t)k = 1 for all
t ∈ [0, L] where L is the total length of the curve. Moreover, (4.10) holds.
We shall show that for every δ > 0 and every T0 ∈ (0, L] there exists a polygonal approximation
m
S
Pδ (T0 ) =
[zk+1 , zk ] of the curve γ([T0 , L]) of accuracy δ > 0. Indeed, let T0 ∈ (0, L] and set
k=0
z0 = γ(T0 ). In view of (4.10) and Lemma 4.3 there exists t− ∈ [T0 − δ, T0 ) such that
γ(t− ) − γ(T0 )
∈ int NΩ(T0 ) (γ(T0 )).
kγ(t− ) − γ(T0 )k
(4.13)
Let T1 > 0 be the infimum of all t− ∈ [T0 − δ, T0 ) such that (4.13) holds, and set z1 = γ(T1 ).
Notice that (z1 − z0 ) ∈ NΩ(T0 ) (γ(T0 )), which yields that for any x ∈ Ω(T0 ), the function
t 7→ d(x, z0 + t(z1 − z0 ))
is increasing. Thus the curve [z1 , z0 ] ∪ {γ(t) : t ≥ T0 } is self-contracted. Since t 7→ γ(t) is the
length parametrization, it follows that kz1 − z0 k ≤ |T1 − T0 | ≤ δ whence
dH ([z1 , z0 ], γ([T1 , T0 ])) ≤ δ.
Notice finally that
[z1 , z0 ] ∪ Ω(T0 ) ⊂ Ω(T1 ).
Repeating the above procedure, we build after m iterations, a decreasing sequence T0 > T1 >
m
S
. . . > Tm+1 and points zk = γ(Tk ), such that
[zk+1 , zk ] is a self-contracted polygonal
k=0
approximation of the curve {γ(t) : t ∈ [Tm+1 , T0 ]} of precision δ > 0.
m
S
Ω(T0 ) ∪
[zk+1 , zk ] ⊂ Ω(Tm+1 ).
Notice also that
k=0
The proof will be complete if we show that with this procedure we reach the initial point t = 0
in a finite number of iterations, i.e. Tm+1 = 0 for some m ≥ 1. Let us assume this is not the case.
Then the aforementioned procedure gives a decreasing sequence {Tm }m & T , where T ≥ 0. Set
z = γ(T ) and zm = γ(Tm ) and notice that {zm } → z. We may assume that {zm }m≥m0 ⊂ B(z, δ).
It follows that for every m ≥ m0 there exists sm > 0 and xm := γ(Tm + sm ) such that
h
x m − zm
z − zm
,
i > 0.
kz − zm k kxm − zm k
(4.14)
(If (4.14) were not true, then we could have taken zm+1 = z and Tm+1 = T a contradiction to
the definition of Tm+1 as an infimum.) Let us now assume that a subsequence of {xm }m remains
away of z. Then taking a converging sub-subsequence {zkm }m and passing to the limit in (4.14)
as m → ∞, we obtain that hγ 0 (T ), ui ≥ 0, for some unit vector u in the cone over Ω(T ) − z,
which contradicts (4.9). It follows that sm → 0 and {xm }m → z.
RECTIFIABILITY OF SELF-CONTRACTED CURVES
IN THE EUCLIDEAN SPACE AND APPLICATIONS 21
For the rest of the proof, let us assume for simplicity that z = 0 and γ 0 (T ) = en = (0, . . . , 1).
(There is no loss of generality in doing so.) Since γ is a C 1 curve, it follows that for any ε > 0,
there exists τ > 0 such that for all t ∈ (T, T + τ ) we have
1 ≥ hen , γ 0 (t)i ≥ 1 − ε
(4.15)
and
γ(t) (4.16)
en −
≤ ε.
kγ(t)k
Pick any m̄ sufficiently large so that Tm̄ +sm̄ < τ, set z̄ = zm̄ = γ(Tm̄ ) and x̄ = xm̄ = γ(Tm̄ +sm̄ ).
Since x̄, z̄ satisfy (4.14) and (4.16) we deduce that
z̄
x̄ − z̄
z̄
h
,
i < 0 and ken −
k < ε,
kz̄k kx̄ − z̄k
kz̄k
x̄−z̄
hence hen , kx̄−z̄k
i < ε, and the last coordinates satisfy
x̄n − z̄ n < εkx̄ − z̄k = ε kγ(Tm̄ + sm̄ ) − γ(Tm̄ )k ≤ εsm̄ .
On the other hand, using (4.15) we deduce
Z sm̄
hen , gamma0 (Tm̄ + s)ids ≥ (1 − ε)sm̄ .
x̄n − z̄ n =
(4.17)
(4.18)
0
Relations (4.17), (4.18) are incompatible for ε < 1/2, leading to a contradiction.
This proves that for some m we get Tm+1 = 0 and zm+1 = γ(0). The proof is complete.
In view of Corollary 4.11 we thus obtain.
Corollary 4.14. Every C 1 convex foliation orbit admits self-contracted polygonal approximations of arbitrary accuracy.
4.3. Convergence of the proximal algorithm. The aim of this section is to provide a different type of application of the notion of self-contracted curves. We recall that for a (nonsmooth)
convex function f : Rn → R which is bounded from below, an initial point x0 ∈ Rn and a
sequence {ti }i ⊂ (0, 1], the algorithm defines a sequence {xi }i≥0 ⊂ Rn called proximal sequence
according to the iteration scheme, see [16, Definition 1.22] for example:
Given xi , define xi+1 as the (unique) solution of the following (strongly convex, coercive)
minimization problem
1
2
min f (x) +
kx − xi k .
(4.19)
x
2ti
A necessary and sufficient optimality condition for the above problem is
0 ∈ ∂f (xi+1 ) + t−1
i (xi+1 − xi )
or equivalently,
xi+1 − xi
∈ −∂f (xi+1 ),
(4.20)
ti
where ∂f is the Fenchel subdifferential of f . Notice that (4.20) can be seen as an implicit
discretization of the subgradient system (1.2) or of the gradient system (1.1), in case f is
smooth.
The proximal algorithm has been introduced in [14]. We refer to [10] for a geometrical
interpretation and to [4] for extensions to nonconvex problems. Let us now recall from [1,
Section 2.4] the following important facts (we give the short proof for completeness):
22
A. DANIILIDIS, G. DAVID, E. DURAND-CARTAGENA, A. LEMENANT
Lemma 4.15 (Geometrical interpretation). Let {xi }i≥0 be a proximal sequence for a convex
function f and set ri := f (xi ). Then:
(i) f (xi+1 ) < f (xi ) if xi+1 6= xi ;
(ii) the point xi+1 is the shortest distance projection of xi on the sublevel set [f ≤ ri+1 ].
Proof. Assertion (i) is obvious, since xi+1 minimizes (4.19), thus for x = xi we get
f (xi+1 ) +
1
kxi+1 − xi k2 ≤ f (xi ).
2ti
Assertion (ii) follows again from (4.19), by considering x ∈ [f ≤ ri+1 ]. Indeed:
f (xi+1 ) +
1
1
1
kxi+1 − xi k2 ≤ f (x) +
kx − xi k2 ≤ ri+1 +
kx − xi k2 .
2ti
2ti
2ti
Since ri+1 = f (xi+1 ) the above yields
kxi+1 − xi k ≤ kx − xi k, for all x ∈ [f ≤ ri+1 ].
The proof is complete.
We obtain the following interesting consequence.
Proposition 4.16 (Self-contractedness of the proximal algorithm). Let {xi }i≥0 be a proximal
sequence for a convex function f . Then the polygonal curve
[
P =
[xi , xi+1 ]
(4.21)
i∈N
is a self-contracted curve.
Proof. Let i1 < i2 < i3 . Notice first that for every x ∈ [f ≤ rm+1 ], the function
t 7→ kxm+1 + t(xm − xm+1 ) − xk
is increasing because the sublevel sets are convex. Now Lemma 4.15 yields that for all i1 ≤ m <
i2 , xi3 ∈ [f ≤ f (xm+1 )] and
kxm+1 − xi3 k ≤ kxm − xi3 k,
which inductively yields
kxi2 − xi3 k ≤ kxi1 − xi3 k.
A simple argument now shows that the polygonal curve is self-contracted.
(4.22)
Combining the above with the length bound given in (2.3) (in view of Remark 3.5) we obtain
the following convergence result.
Theorem 4.17 (Convergence of the proximal algorithm). Let f : Rn → R be convex and
bounded from below. Let x0 ∈ Rn be an initial point for the proximal algorithm with parameters
{ti }i ⊂ (0, 1]. Then the proximal sequence {xi }i converges to some point x∞ ∈ Rn with a fast
rate of convergence. That is, there exists c > 0, depending only on the dimension, such that
X
kxi − xi+1 k ≤ ckx0 − x∞ k.
i∈N
RECTIFIABILITY OF SELF-CONTRACTED CURVES
IN THE EUCLIDEAN SPACE AND APPLICATIONS 23
Proof. By Proposition 4.16 the polygonal curve P defined in (4.21) is self-contracted. Consequently, by Remark 3.5 we know that (2.3) applies, yielding
X
`(P ) =
kxi − xi+1 k ≤ Cn diam K,
n∈N
where K is any compact convex set containing {xi }i . In addition, letting i1 = 0 and i3 → +∞
in (4.22) we get
kxi − x∞ k ≤ kx0 − x∞ k,
for all i ∈ N, which implies that {xi }i ⊂ B(x∞ , R) with R = kx0 − x∞ k and the result follows
for c = 2Cn .
Theorem 4.17 guarantees the fast convergence of the proximal sequence {xi }i towards a limit
point, with a convergence rate independent of the choice of the parameters {ti }i . Notice however
that for an arbitrary choice of parameters, the limit point might not be a critical point of f as
shows the following example:
Example 4.18 (Convergence to a noncritical point). Consider the (coercive, C 1 ) convex function f : R → R defined by
x2 ,
|x| ≤ 1/2
f (x) =
1
|x| − 4 , |x| ≥ 1/2
Then the proximal algorithm {xi }i , initialized at the point x0 = 2 and corresponding to the
parameters ti = 1/2i+1 , i ∈ N, converges to the noncritical point x̄ = 1.
In the aboveP
example, the choice of the parameters {ti } in the proximal algorithm has the
drawback that i ti < +∞. In practice the choice of the step-size parameters {ti }i is crucial to
obtain the convergence of the sequence
{f (xi )}i towards a critical value; a standard choice for
P
this is any sequence satisfying
ti = +∞, see for instance [8].
References
[1] Bolte, J., Daniilidis, A., Ley, O., Mazet, L., Characterizations of Lojasiewicz inequalities: subgradient
flows, talweg, convexity, Trans. Amer. Math. Soc. 362 (2010), 3319–3363.
[2] Brézis, H., Opérateurs maximaux monotones et semi-groupes de contractions dans les espaces de Hilbert
(French), North-Holland Mathematics Studies 5, (North-Holland Publishing Co., 1973).
[3] Burago, D., Burago, Y., Ivanov, S., A course in metric geometry, Graduate Studies in Mathematics,
Vol. 33, American Mathematical Society, Providence, RI, 2001.
[4] Combettes, P. & Pennanen, T., Proximal methods for cohypomonotone operators, SIAM J. Control
Optim. 43 (2004), 731–742.
[5] Daniilidis, A., Garcia Ramos, Y., Some remarks on the class of continuous (semi-)strictly quasiconvex
functions, J. Optim. Theory Appl. 133 (2007), 37–48.
[6] Daniilidis, A. Ley, O., Sabourau, S., Asymptotic behaviour of self-contracted planar curves and gradient
orbits of convex functions, J. Math. Pures Appl. 94 (2010), 183–199.
[7] Giannotti, C. Spiro, A., Steepest descent curves of convex functions on surfaces of constant curvature,
Isr. J. Math. 191 (2012), 279–306.
[8] Güler, O. On the convergence of the proximal point algorithm for convex minimization, SIAM J. Control
Optim. 29 (1991), 403-419.
[9] Kurdyka, K., On gradients of functions definable in o-minimal structures, Ann. Inst. Fourier 48 (1998),
769-783.
[10] Lemaire, B. About the convergence of the proximal method. Advances in optimization (Lambrecht, 1991),
3951, Lecture Notes in Econom. and Math. Systems 382, Springer, Berlin, 1992.
[11] Lojasiewicz, S., “Une propriété topologique des sous-ensembles analytiques réels.”, in: Les Équations aux
Dérivées Partielles, pp. 87–89, Éditions du centre National de la Recherche Scientifique, Paris, 1963.
24
A. DANIILIDIS, G. DAVID, E. DURAND-CARTAGENA, A. LEMENANT
[12] Manselli, P. Pucci, C., Uniqueness results for evolutes and self-evolvents, Boll. Un. Mat. Ital. A 5 (1991),
373–379.
[13] Manselli, P. Pucci, C., Maximum length of steepest descent curves for quasi-convex functions, Geom.
Dedicata 38 (1991), 211–227.
[14] Martinet, B., Régularisation d’inéquations variationnelles par approximations successives. (French) Rev.
Française Informat. Recherche Opérationnelle, 4, Ser. R3 (1970) 154-158.
[15] Palis, J. & De Melo, W., Geometric theory of dynamical systems. An introduction, (Translated from the
Portuguese by A. K. Manning), Springer-Verlag, New York-Berlin, 1982.
[16] Rockafellar, R.T. & Wets, R., Variational Analysis, Grundlehren der Mathematischen, Wissenschaften,
Vol. 317, (Springer, 1998).
————————————————————Aris Daniilidis
Departament de Matemàtiques, C1/308
Universitat Autònoma de Barcelona
E-08193 Bellaterra, Spain
E-mail: [email protected]
http://mat.uab.es/~arisd
Research supported by the grant MTM2011-29064-C01 (Spain).
Guy David
Laboratoire de mathmatiques, UMR 8628, Bâtiment 425
Université Paris-Sud, F-91405 Orsay Cedex, France
et Institut Universitaire de France
E-mail: [email protected]
http://www.math.u-psud.fr/~gdavid/
Estibalitz Durand-Cartagena
Departamento de Matemática Aplicada
ETSI Industriales, UNED
Juan del Rosal 12, Ciudad Universitaria, E-28040 Madrid, Spain
E-mail: [email protected]
https://dl.dropbox.com/u/20498057/Esti/Home Page.html
Antoine Lemenant
Laboratoire Jacques Louis Lions (CNRS UMR 7598)
Université Paris 7 (Denis Diderot)
175, rue du Chevaleret, F-75013 Paris, France
E-mail: [email protected]
http://www.ann.jussieu.fr/~lemenant/